mister_monster on Nostr: The study concludes that all of the cutting edge big tech models show capability to ...
The study concludes that all of the cutting edge big tech models show capability to engage in dishonesty. That leads me to believe there are probably problems with their methodology. The way these models work I don't think they're architecturally capable of meta thought, or thought at all really but if we call what they're doing "thought". I'd be willing to bet the model was asked to put some stuff in the thinking tags, and it's just doing that. I'd also be willing to bet the model was trained on articles and literature about dystopian AI outcomes.
Published at
2025-01-31 07:06:50Event JSON
{
"id": "0ed30e58f417599340f4d122bdb69b766e33753f1eff3a1218391f7a186359d1",
"pubkey": "dd2057556f88a64cacd075d007f1be480f949c91fd6d0c4d593baccdb2aabde2",
"created_at": 1738307210,
"kind": 1,
"tags": [
[
"e",
"49401497c360a46f0b868d203d8fd53b3fa4b6ed4126bc7be1e57cad169792f5",
"",
"root"
],
[
"p",
"55f573b651eff351db57b0601d23022d8c532f9825db10a5733ebf39be4aa21b"
]
],
"content": "The study concludes that all of the cutting edge big tech models show capability to engage in dishonesty. That leads me to believe there are probably problems with their methodology. The way these models work I don't think they're architecturally capable of meta thought, or thought at all really but if we call what they're doing \"thought\". I'd be willing to bet the model was asked to put some stuff in the thinking tags, and it's just doing that. I'd also be willing to bet the model was trained on articles and literature about dystopian AI outcomes.",
"sig": "f99d7b48c87e5ad8839e025e18154f7d377183244ed00bf899bb0dda931d1dfcfa79a0becb3017b2beabe260446445d43141b313ed16ef13d1705343348dde08"
}
